Although compile-time optimizations generally improve program performance, degradations caused by individual techniques are to be expected. Feedback-directed optimizations have recently begun to address this issue, by factoring runtime information into the decision process of which compiler optimization to apply where and when. While improvements for small sets of optimization techniques have been demonstrated, little empirical knowledge exists on the performance behavior of the large number of today's optimization techniques. This is especially true for the interaction of such techniques, which we have found to be of significant importance in navigating the search space of the best combination of techniques. The contribution of this paper is in (1) providing such empirical knowledge and (2) developing algorithms for efficiently navigating and pruning the search space. To this end, we evaluate the optimization techniques of GCC on both a Pentium IV machine and a SPARC II machine, by measuring the performance of the SPEC CPU2000 benchmarks under different compiler ags. We analyze the performance losses that result from individual optimizations. We then present three heuristic algorithms that search for the best combination of compiler techniques using measured runtime as feedback.
Date of this Version